Abstract: The silicon content in hot metal is a key performance indicator for evaluating both the quality of the hot metal and the operational state of the blast furnace. However, due to the complex nonlinear relationships and time-lag effects between the silicon content and process variables, the control of hot metal quality remains largely empirical. To address this, a prediction method for hot metal silicon content is proposed, combining TimeGAN-based data augmentation with the TransBiGRU-AKSM model. First, to tackle the issue of limited labeled data, a time-series generative adversarial network is employed to effectively expand the dataset and enhance the generalization ability of the model. Second, to efficiently capture long-term dependencies in time series and improve the accuracy in tracking silicon content trends, the model integrates the global feature extraction capability of the Transformer with the bidirectional temporal learning strength of BiGRU. Moreover, to enhance adaptability under complex blast furnace conditions, an Adaptive Key-feature Selection Mechanism is introduced, enabling dynamic selection and weighted processing of key features. Finally, the effectiveness and accuracy of the proposed method are validated using real industrial data.